director-qa

Director — Gate

Run QA gates on each generated clip. Three gates in order — anatomy first (cheapest), then critic, then speech (if enabled). If any gate rejects, regenerate the clip via /director-clip-gen and re-run gates. Max 5 retries per scene. Abort if all fail.

Gate 1 — Anatomy

Extract 3 keyframes (start, middle, near-end) from the clip. Upload each via POST /v1/upload. Run spatial check on each:

POST /v1/watch
{
  "mediaUrl": "<frameUrl>",
  "prompt": "Spatial anatomy check of AI-generated video frame. Intended scene: '<scene.prompt>'. Trace every visible arm from fingertips to shoulder — does it connect to a torso? Check for: limb-to-torso connectivity, spatial consistency, disembodied body parts, anatomically possible limb count, correct joint positions. Score 6+ ONLY if all visible body parts are spatially connected and anatomically plausible."
}

Auto-reject if any frame scores below 8.

Gate 2 — Critic

Upload the full clip via POST /v1/upload. Run visual QA:

POST /v1/watch
{
  "mediaUrl": "<clipUrl>",
  "prompt": "AI-generated video clip. Intended scene: '<scene.prompt>'. Check: face consistency, spatial anatomy, visual artifacts, motion quality, audio sync (if applicable). Score 1-10. List issues with severity. Flag CRITICAL: morphing, warping, structural drift, broken physics, disconnected anatomy. Ignore text rendering on clothing."
}

Auto-reject if:

• Score < 8/10
• Any major/critical issues (excluding text_rendering category)

Adapt the critic prompt to the content style — if b-roll, reward natural motion and penalize stiffness. If talking head, check for stillness and eye contact.

Gate 3 — Speech (if enabled)

Skip this gate if the video has no spoken dialogue.

POST /v1/transcribe
{
  "mediaUrl": "<clipUrl>"
}

Compare transcript word-by-word against the intended dialogue for the scene.

Auto-reject if:

• Missing or wrong words
• Repeated phrases (same 3+ word sequence appears consecutively)
• Audio events during speech (laughs, sighs, clicks)

Non-blocking:

• Numbers normalized — "8" matches "eight", "12" matches "twelve"
• Peripheral noise before first word or after last word — trimmed in post

Retry Logic

If any gate rejects a clip:

1. Log which gate failed and why
2. Regenerate the clip via /director-clip-gen (same scene, same frames)
3. Re-run all gates from the beginning
4. Max 5 total attempts per scene
5. If attempt 5 fails, abort the entire pipeline — don't continue to post-production with missing scenes

Output

For each scene that passes all gates:

• Clip URL
• Critic score
• Number of attempts
• Last frame URL (via POST /v1/extract with frames: "last") for frame chaining to the next scene

Pass all passed clips to /director-post.